view online

Welcome to AiTuts, the only newsletter that encourages you to gather a crew of 9 friends, sail to the end of the high seas, and seek out the greatest treasure in the world.

After you've learned about AI, of course.
​
We're going with a new format today:

✨3 Things to Try

1/ The Era of SDXL v1.0

SDXL v1.0 has been out for 10 days now.

Just enough time for the jury to reach a verdict:

SDXL images are on-par with (or even better than) Midjourney images, if you don't just use the base model.

SDXL has 2 models and a 2 step process. The base model generates images and the refiner model makes them more detailed.

The results are incredible.

StabilityAI reports that users prefer SDXL v1.0 results (base + refiner) over commercial alternatives in a blind test (real or flagrant self-promotion? that's for you to decide).

That's not to forget: the avalanche of coming SDXL custom models.

Remember how there are a gazillion Stable Diffusion models?

Most of these are trained on top of Stable Diffusion v1.5, and improve the quality of the original considerably.

When the custom models start rolling in for SDXL, the images will get even better. The SDXL era has begun.

2/ ComfyUI is not comfy, but that's not stopping anyone

What is ComfyUI and why are people moving over from AUTOMATIC1111?

For the new folks: AUTOMATIC1111 is the most popular tool for running Stable Diffusion on your own computer.

But after the release of SDXL, people have been moving over to a tool called ComfyUI in droves.

ComfyUI looks like this:

I know, I know.

The memes write themselves:

But did you know that the StabilityAI team, creators of Stable Diffusion, use ComfyUI to test Stable Diffusion?

And that they hired the anonymous creator to help them develop in-house tools?

So what’s the big deal?

• ComfyUI is optimized to run on GPUs, so it generates much faster.
• Instead of doing your generation process in multiple steps (generate, img2img, upscale, extensions like ControlNet), you run everything with one click.

ComfyUI just might become the professional standard for Stable Diffusion. Now if it were just a bit easier to learn....

3/ Run Llama-2 on your computer

The LLaMA large language model release in March of this year kicked off a cambrian explosion of open source models.

Llama-2 is the follow up.

A fine-tuned version of Llama-2, Llama-2-70b-instruct-v2, currently sits at the top of the leaderboards of Open Source Large Language Models.

In fact, 7 of the top 10 models are fine-tunes of Llama-2!

Models are scored by both humans and AI, based on their performance on tasks such as brainstorming, creative generation, reasoning, question answering, summarization, and code generation.

The future of large language models looks to be split between closed and open source:

• Companies who don't have machine learning engineers, or don't want the hassle will pay providers like OpenAI to use models like GPT-4.
• Companies who want privacy or customizability will fine-tuning their own models on top of open source models.

Here's a guide to get you set up with Llama-2:

⛵ Free Handbook: Midjourney for Fantasy Art

Midjourney for Fantasy Art is a book for worldbuilders. It'll teach you prompts and techniques for fantasy heroes and lost worlds for games, comics & films.

Yeah, yeah, it's a book for V5.1... but the prompts look great in V5.2 as well (we promise!)

✨ 3 Roundups

1/ Creative Roundup

Midjourney V5.3 is coming soon, says CEO David Holz. Midjourney V6 has been postponed for further improvements. The team is working on building generation features into the actual Midjourney website. Midjourney inpainting is ready to release, but Discord will need to update their UI for it to work [link]

Meta releases AudioCraft, an open source suite of AI tools for creating music from text. [link]

2/ Chatbot Roundup

OpenAI rolled out a bunch of features this week to make ChatGPT more user friendly: prompt examples, suggested replies and keyboard shortcuts [link]

Agnaistic is a free-to-use, "bring your own chatbot" tool with a slick interface and plenty of features. You hook up a services like OpenAI, NovelAI, GooseAI, Scale, Claude, and get chatting [link]

3/ In the News

Facebook is preparing to launch a range of AI-powered chatbots that exhibit different personalities and talk to users [link]

Apple App Store in China removes numerous AI apps including OpenCat, a popular ChatGPT client, ahead of new Chinese AI regulations. [link]

IBM and NASA launch a geospatial foundational model on HuggingFace. The model operates on satellite images, and can perform tasks like flood prediction and crop classification. [link]

That's a wrap!

You can reply to this email directly. We read every reply.

Till next time,

Yubin & Crew

1223 Cleveland Ave #200, San Diego, CA 92103
​Unsubscribe · Preferences​